---
title: Energy-Based ADAPT-VQE Overview
url: https://www.emergentmind.com/topics/energy-based-adapt-vqe
type: topic
---

# Energy-Based ADAPT-VQE Overview

Energy-based ADAPT-VQE designates adaptive variational quantum eigensolver schemes in which the next ansatz generator is ranked by an explicit estimate of variational energy lowering, rather than solely by a zero-parameter commutator or gradient magnitude. In the stricter usage that emerged later, this distinguishes energy-based selection from the original ADAPT-VQE rule \( \left.\partial E/\partial \theta_i\right|_{\theta_i=0}=\langle \psi|[H,\tau_i]|\psi\rangle \), even though both classes minimize the same ground-state objective \(E(\boldsymbol{\theta})=\langle \psi(\boldsymbol{\theta})|\hat H|\psi(\boldsymbol{\theta})\rangle\) [2011.01279]. The literature is not fully uniform, however: some works describe ADAPT-VQE as “energy-based” in the broader sense that the algorithm is organized around energy minimization, while their adaptive screening remains gradient-driven [2109.12126]. This taxonomic ambiguity is central to the subject, because many practically important developments around ADAPT-VQE—batching, RDM reconstruction, informationally complete measurements, pruning, and Hamiltonian-aware proxies—improve adaptive VQE without becoming strict energy-based operator selection [2108.11167].

## 1. Definition and taxonomic scope

Under the strict definition, an energy-based ADAPT-VQE iteration evaluates, for every candidate generator \(\hat\tau_g\) in the pool, a one-parameter trial landscape
\[
E_g^{(n+1)}(\theta_g)=\bra{\Psi^{(n)}(\boldsymbol{\theta})} e^{-\theta_g \hat\tau_g}\hat H e^{\theta_g \hat\tau_g} \ket{\Psi^{(n)}(\boldsymbol{\theta})},
\]
and appends the operator whose minimized landscape yields the lowest energy [2606.04786]. This differs from original ADAPT-VQE, which ranks candidates by the magnitude of the derivative at \(\theta_g=0\),
\[
\left.\frac{\partial E^{(n)}}{\partial \theta_g}\right|_{\theta_g=0}
=
\bra{\Psi^{(n)}(\boldsymbol{\theta})}[\hat H,\hat\tau_g]\ket{\Psi^{(n)}(\boldsymbol{\theta})},
\]
then reoptimizes after appending the largest-gradient generator [2011.01279].

The distinction matters because gradient selection is a first-order local criterion, whereas strict energy-based selection evaluates an actual one-dimensional variational benefit. Several later papers treat this distinction as substantive: some explicitly note that earlier ADAPT studies remain gradient-based even when they benchmark energy minimization or reduce energy-measurement overhead [2108.11167], and parameter-based work cites prior movement from gradients to energy-based criteria precisely because large gradients can correspond to redundant operators [2602.04253].

| Variant | Selection rule | Representative paper |
|---|---|---|
| Gradient-based ADAPT-VQE | Largest \(\left|\langle \psi|[H,\tau_i]|\psi\rangle\right|\) | [2011.01279] |
| Strict energy-based ADAPT-VQE | Lowest \(\min_{\theta_i} E_i(\theta_i)\) | [2606.04786] |
| Energy-reduction excited-state ADAPT | Largest explicit \(\Delta E_p^{[m]}\) under deflated Hamiltonian | [2106.06296] |
| Hamiltonian-aware proxy ADAPT | Largest \(|h_i\sin(2\theta_i^*)|\) | [2606.13118] |

This suggests that “energy-based ADAPT-VQE” is best treated as a family of objective-aware operator-selection rules, with direct energy-drop ranking at one pole and lower-cost surrogates at the other.

## 2. Canonical algorithmic structure

In fermionic ADAPT-VQE, the ansatz after \(n\) adaptive steps is typically written as
\[
\ket{\Psi^{(n)}(\boldsymbol{\theta})}
=
\hat U^{(n)}(\theta^{(n)})\cdots \hat U^{(1)}(\theta^{(1)})\ket{\Psi_{\mathrm{ref}}},
\qquad
\hat U^{(n)}(\theta^{(n)})=e^{\theta^{(n)}\hat\tau_g},
\]
with \(\ket{\Psi_{\mathrm{ref}}}\) the Hartree–Fock determinant in the fixed-orbital setting [2606.04786]. The adaptive loop has three separable components: operator-pool screening, ansatz growth, and post-growth parameter optimization. Energy-based ADAPT-VQE modifies only the first of these in principle, but in practice it also changes initialization because the winning one-parameter landscape supplies a nonzero warm start for the appended operator [2606.04786].

For the direct fermionic formulation, the pool is built from anti-Hermitian singles and doubles. In orbital-optimized variants, orbital rotations
\[
\hat U(\boldsymbol{\kappa}) = \exp\!\left\{ \sum_{p>q}\kappa_{pq}(\hat E_{pq}-\hat E_{qp}) \right\}
\]
move explicit singles into the orbital manifold, so the adaptive pool may be reduced to doubles only [2606.04786]. This is not merely a technical refinement: the 2026 benchmark found that as correlation increases, full ansatz reoptimization and orbital optimization become the main factors governing convergence and total resource cost, often outweighing the difference between gradient-based and energy-based selection [2606.04786].

A parallel excited-state construction appears in e-QEB-ADAPT-VQE. There the effective Hamiltonian is modified by overlap penalties,
\[
H_k = H + \sum_{r=0}^{k-1}\alpha_r |E_r\rangle\langle E_r|,
\]
and candidate qubit-excitation evolutions are ranked by explicit energy reduction under \(H_k\), first through one-parameter screening and then through full reoptimization of a shortlist [2106.06296]. In that sense, excited-state energy-based ADAPT is not a separate idea but a direct extension of the same operator-selection philosophy to deflated objectives.

## 3. Direct realizations of energy-based selection

A clear early direct realization is the parabolic-optimization framework of “Efficient Parabolic Optimisation Algorithm for adaptive VQE implementations” [2110.12756]. There, for each candidate excitation \(\hat A_i\), a one-dimensional parabolic optimizer estimates the minimized energy
\[
E_i=\langle\Psi'(x_i)|\hat H|\Psi'(x_i)\rangle,
\qquad
|\Psi'(x_i)\rangle=\hat A_i(x_i)|\Psi\rangle,
\]
and the algorithm chooses the operator yielding the lowest energy. Global reoptimization is then carried out with an \(n\)-dimensional parabolic optimizer. In noiseless statevector simulations on HF, LiH, H\(_2\)O, and BeH\(_2\), the parabolic optimizer outperformed Nelder–Mead in the number of experiments required to reach a given energy accuracy, with no significant impact on the number of CNOTs required to reach that accuracy [2110.12756]. This formulation is important historically because it shows energy-based operator choice and accelerated reoptimization appearing together rather than as separate design problems.

The most explicit excited-state realization is e-QEB-ADAPT-VQE [2106.06296]. For each candidate qubit excitation evolution \(\tilde A_p(\theta_p)\), the method computes
\[
\min_{\theta_p}
\langle \psi^{[m-1]}|
\tilde A_p^\dagger(\theta_p)\,H_k\,\tilde A_p(\theta_p)
|\psi^{[m-1]}\rangle,
\]
uses the corresponding one-parameter energy reduction to shortlist operators, and then computes
\[
\Delta E_p^{[m]}
=
E^{[m-1]}
-
\min_{\boldsymbol{\theta}^{[m-1]},\theta_p}
E(\boldsymbol{\theta}^{[m-1]},\theta_p)
\]
for the shortlisted set. The final appended operator is the one with maximal \(\Delta E_p^{[m]}\). On LiH and BeH\(_2\), this direct energy-reduction criterion produced first-excited-state circuits with maximum CNOT counts of 311 and 896, respectively, versus 3496 and 8980 for UCCSD, and 29447 and 64064 for GUCCSD [2106.06296]. The same paper also records an important caveat: greedy energy-based excited-state growth can converge to the wrong member of a near-degenerate manifold when a nearby less correlated state offers larger immediate energy reductions [2106.06296].

The most systematic modern benchmark of strict energy-based fermionic selection is “Resource-efficient energy-based operator selection in fermionic ADAPT-VQE via exact Hamiltonian transformation” [2606.04786]. That work reformulates the one-parameter landscape through a generator-dependent fragmentation
\[
\hat H=\hat H_{g,0}+\hat H_{g,1}+\hat H_{g,4},
\]
obtained from an exact transformed-Hamiltonian analysis. The resulting method is mathematically identical to standard fermionic Rotoselect: it returns the same one-parameter energy landscape, the same minimizing angle, and therefore the same ansatz. Its contribution is resource reduction, not a changed scoring rule [2606.04786].

## 4. Measurement overhead and optimization tradeoffs

The principal historical objection to energy-based ADAPT-VQE is measurement cost. Parameter-based work states that earlier energy-based criteria “incur a quartic additional measurement cost,” which is precisely why alternative ranking surrogates were proposed [2602.04253]. The 2026 exact-transformation benchmark showed that this cost can be reduced substantially without abandoning strict energy-based selection. Relative selection-cost ratios were reported as follows [2606.04786]:

| System | RS/RSe | RSe/GB |
|---|---:|---:|
| LiH (eq.) | 1.97 | 1.09 |
| BeH\(_2\) (eq.) | 2.06 | 1.06 |
| H\(_2\)O (eq.) | 2.11 | 1.17 |

Here RS is standard fermionic Rotoselect, RSe is the exact transformed-Hamiltonian implementation, and GB is gradient-based ADAPT screening. The paper’s central empirical conclusion is that exact Hamiltonian transformations reduce standard fermionic Rotoselect cost by about a factor of two, bringing energy-based selection close to gradient-based ADAPT-VQE [2606.04786].

The same benchmark also clarifies when strict energy-based screening matters most. In equilibrium LiH, the `last` strategy—optimizing only the appended operator—combined with energy-based selection can construct an accurate ansatz while avoiding any VQE optimization for the new layer [2606.04786]. In more correlated equilibrium BeH\(_2\) and H\(_2\)O, and especially at stretched geometries, `last` strategies stagnate; full reoptimization and orbital optimization dominate performance. A plausible implication is that direct energy-based operator scoring improves local step quality, but it does not remove the need for global variational relaxation once correlation becomes distributed over many adaptive layers.

Several adjacent developments are not strict energy-based selection but are directly relevant to its practicality. Classical pre-optimization with the sparse wavefunction circuit solver (SWCS) embeds standard derivative-based ADAPT-VQE in a tunable high-performance-computing workflow and pushes molecular simulations to systems with up to 52 spin-orbitals, including C\(_2\)/cc-pVDZ with 27,944,940 symmetry-allowed determinants [2411.07920]. AIM-ADAPT-VQE uses informationally complete POVMs so that the same dataset used for energy estimation can also estimate all commutators, giving no additional measurement overhead for operator selection in the tested H\(_4\) and N\(_2\) cases [2212.09719]. RDM-based ADAPT-V and ADAPT-Vx rewrite residual-gradient evaluation in terms of low-order reduced density matrices, reducing screening from \(\mathcal O(N^8)\) to \(\mathcal O(N^4)\) in the approximate Valdemoro-reconstructed variant [2012.07047]. These works do not validate strict energy-based selection directly, but they define the resource landscape within which it must compete.

## 5. Excited states, overlap guidance, and non-variational relatives

Energy-based adaptive selection generalizes naturally to excited states. In e-QEB-ADAPT-VQE, the modified Hamiltonian \(H_k\) embeds lower-state projectors, and the adaptive rule selects operators by explicit energy reduction under that deflated objective [2106.06296]. A related but distinct path appears in spin-restricted ADAPT-VQD, where the selection rule remains gradient-based but the objective is the deflated VQD functional
\[
F(\boldsymbol{\theta})=
\langle\Psi(\boldsymbol{\theta})|\hat H|\Psi(\boldsymbol{\theta})\rangle
+
\sum_i \beta_i |\langle \Phi_i|\Psi(\boldsymbol{\theta})\rangle|^2.
\]
That work produced the most compact LiH excited-state circuits among the compared VQD ansätze, but it should be classified as gradient-based under an energy-plus-deflation objective rather than direct energy-based operator ranking [2105.10275].

Overlap-ADAPT-VQE reframes the issue from another direction. The method replaces local energy descent during ansatz growth by overlap growth toward an intermediate correlated target state, then uses the resulting compact ansatz as initialization for a new standard ADAPT procedure [2301.10196]. The paper’s interpretation is that standard energy-based ADAPT-VQE can become trapped on difficult regions of the energy landscape, especially in strongly correlated systems, producing over-parameterized ansätze. In stretched H\(_6\), CIPSI-Overlap-ADAPT-VQE reached chemical accuracy with only 40 parameters, whereas prior QEB-ADAPT-VQE results required more than 150 parameters [2301.10196]. This suggests that energy-based ADAPT-VQE is not only a selection rule but also part of a broader ansatz-discovery workflow that can benefit from preconditioning.

NoVa-ADAPT occupies a neighboring, non-variational corner of the design space [2411.09736]. It keeps the ADAPT operator-selection rule
\[
\left.\frac{\partial E_{A_i}}{\partial \theta}\right|_{\theta=0}
=
i\langle \psi^{(n-1)}|[H,A_i]|\psi^{(n-1)}\rangle
\]
but replaces post-growth VQE reoptimization by a direct gradient-based parameter update
\[
\eta_n = -\gamma
\left.\frac{\partial E_{A^{(n)}}}{\partial \theta}\right|_{\theta=0}.
\]
This is not energy-based ADAPT-VQE in the strict sense, yet it is highly relevant because it demonstrates that adaptive state preparation can use energy information not only for operator choice but also for parameter assignment. On H\(_4\), NoVa-ADAPT reached chemical accuracy at a measurement cost similar to ADAPT-VQE, despite using more operators [2411.09736].

## 6. Misconceptions, limitations, and current outlook

A recurring misconception is that any ADAPT-VQE study centered on energy minimization is “energy-based.” Multiple benchmark and application papers remain gradient-selected even when they focus on energies, reaction energetics, or measurement reduction. Batched ADAPT-VQE for O\(_2\), CO, and CO\(_2\) is a prominent example: the operator choice is still gradient-ranked, only with multiple operators added when their gradients are close to the largest [2108.11167]. Likewise, standard molecular benchmarks on H\(_2\), NaH, and KH study how well ADAPT minimizes the energy objective, but operator screening remains the original commutator rule [2011.01279]. The strict definition therefore remains essential.

A second limitation is that operator ranking alone does not determine performance. Gradient-based ADAPT-VQE was shown to be robust to rough parameter landscapes chiefly because of one-operator-at-a-time growth and warm-started optimization, not because local minima vanish; even when one step falls into a trap, ansatz growth can “burrow” toward the solution [2204.07179]. However, the same study notes that the strongest barren-plateau-avoidance argument is specific to gradient-selected ADAPT, because it relies on the newly appended parameter being guaranteed to have a resolvable nonzero gradient [2204.07179]. A plausible implication is that strict energy-based ADAPT-VQE inherits the warm-start benefit but does not automatically inherit the same gradient-threshold guarantee.

A third limitation concerns stagnation mechanisms other than operator misranking. “Gradient troughs” arise when appending-position gradients become very small even though the minimum energy has not been reached, and the 2025 insertion-policy study attributes them partly to repeated use of the same ansatz location for new operators [2512.25004]. Since insertion position is largely orthogonal to the ranking metric, this lesson should transfer to energy-based ADAPT-VQE: appending-only energy scoring may also miss productive operator-position moves.

Current low-cost alternatives underscore the same theme. Param-ADAPT-VQE replaces gradient magnitude by the locally optimized parameter magnitude \(|\theta_i^*|\), together with sub-Hamiltonian screening and hot-start reoptimization [2602.04253]. HA-ADAPT-VQE uses the Hamiltonian-aware proxy \(|h_i\sin(2\theta_i^*)|\), explicitly stating that this is not the true energetic contribution but a low-cost approximation [2606.13118]. Pruned-ADAPT-VQE shows that even after adaptive growth, redundant operators can remain because of poor selection, operator reordering, or fading, and removes them through a cost-free post-optimization refinement rule [2504.04652]. These methods collectively suggest that strict energy-based ADAPT-VQE is best viewed not as the only “objective-aware” strategy, but as the reference point against which increasingly resource-conscious surrogates are calibrated.

The contemporary picture is therefore mixed but clearer than before. Direct energy-based operator selection is now technically viable in fermionic ADAPT-VQE at a cost close to gradient-based screening when exact Hamiltonian transformations are used [2606.04786]. In weakly correlated regimes, it can enable efficient ansatz growth with little or no extra VQE optimization [2606.04786]. In more strongly correlated regimes, full reoptimization, orbital optimization, insertion policy, and ansatz compaction become at least as important as the scoring rule itself [2512.25004]. Energy-based ADAPT-VQE is consequently best understood as one axis in a broader adaptive-design space: direct energy lowering provides the most literal alignment with the variational objective, but practical performance depends on how that criterion is embedded in measurement architecture, optimization strategy, orbital treatment, and ansatz-management heuristics.

Source: https://www.emergentmind.com/topics/energy-based-adapt-vqe